Individual Tree Detection Methods Research Articles

Mangrove individual tree detection based on the uncrewed aerial vehicle multispectral imagery

Mangroves are one of the most threatened and vulnerable ecosystems in the world. They dominate the coastal wetlands of tropical and subtropical regions. These forest areas are frequently difficult to access, typical methods of forest inventory are difficult undertakings. Thus, the acquisition of forest attributes through remote sensing, such as the validated individual tree detection (ITD) aim can support management and conservation proposals. This study comprehensively validated ITD from high-resolution multispectral uncrewed aerial vehicle (UAV) surveys, forest inventory data, multiresolution segmentation, mean shift segmentation and gaussian model the aim of quantifying the canopies in different areas of mangrove. The validations methods were based on the number plot size, the polygon canopy, and point canopy obtained according to simple sampling plan statistics. The results demonstrated that when using plot size inside the flight area to calibrate the ITD methods, the gaussian model and the segmentation methods can reach an acceptable quality limit (AQL) of 15%. On the other hand, this study suggests as alternative to ITD a tree like objects (TLO) method for flying through dense forests without forest inventory with AQL of 40%. This study is critical for conservation and management efforts because it applies the reliability of multispectral UAV surveys to assessing tree abundance in mangrove forests and can be applied to quantify biomass and blue carbon in dense forests.

A New Strategy for Individual Tree Detection and Segmentation from Leaf-on and Leaf-off UAV-LiDAR Point Clouds Based on Automatic Detection of Seed Points

Accurate and efficient estimation of forest volume or biomass is critical for carbon cycles, forest management, and the timber industry. Individual tree detection and segmentation (ITDS) is the first and key step to ensure the accurate extraction of detailed forest structure parameters from LiDAR (light detection and ranging). However, ITDS is still a challenge to achieve using UAV-LiDAR (LiDAR from Unmanned Aerial Vehicles) in broadleaved forests due to the irregular and overlapped canopies. We developed an efficient and accurate ITDS framework for broadleaved forests based on UAV-LiDAR point clouds. It involves ITD (individual tree detection) from point clouds taken during the leaf-off season, initial ITS (individual tree segmentation) based on the seed points from ITD, and improvement of initial ITS through a refining process. The results indicate that this new proposed strategy efficiently provides accurate results for ITDS. We show the following: (1) point-cloud-based ITD methods, especially the Mean Shift, perform better for seed point selection than CHM-based (Canopy Height Model) ITD methods on the point clouds from leaf-off seasons; (2) seed points significantly improved the accuracy and efficiency of ITS algorithms; (3) the refining process using DBSCAN (density-based spatial clustering of applications with noise) and kNN (k-Nearest Neighbor classifier) classification significantly reduced edge errors in ITS results. Our study developed a novel ITDS strategy for UAV-LiDAR point clouds that demonstrates proficiency in dense deciduous broadleaved forests, and this proposed ITDS framework could be applied to single-phase point clouds instead of the multi-temporal LiDAR data in the future if the point clouds have detailed tree trunk points.

Multi-level self-adaptive individual tree detection for coniferous forest using airborne LiDAR

• Advantages of both CHM-based and point-based methods are combined. • Crown width scales are estimated automatically based on gradient magnitudes. • Multi-scales segmented results are fused and optimized gradually. • Neighboring trees are separated based on probability density function splitting. To obtain satisfying results of individual tree detection from LiDAR points, parameters using traditional methods usually need to be adjusted by trials and errors. When encountering complex forest environments, the detection accuracy cannot be satisfied. To resolve this, a multi-level self-adaptive individual tree detection method was presented in this paper. The proposed method can be seen as a hybrid model, which combined the strength of both raster-based and point-based methods. Raster-based strategy was first used for achieving initial trees detection results, while the point-based strategy was adopted for optimizing the clustered trees. In the proposed method, crown width scales were estimated automatically. Meanwhile, multi-scales segmented results were fused together to take advantage of segmented results of both larger and small scales. Six different coniferous forests were adopted for testing. Experimental result shows that this study achieved the lowest omission and commission errors comparing with other three classical approaches. Meanwhile, the average F1 score in this paper is 0.84, which is much highest out of other methods.

Cross-Comparison of Individual Tree Detection Methods Using Low and High Pulse Density Airborne Laser Scanning Data

Numerous individual tree detection (ITD) methods have been developed for use with airborne laser scanning (ALS) data to provide tree-scale forest inventories across large spatial extents. Despite the growing number of methods, relatively few have been comparatively assessed using a single benchmark forest inventory validation dataset, limiting their operational application. In this study, we assessed seven ITD methods, representing three common approaches (point-cloud-based, raster-based, hybrid), across coniferous forest stands with diverse structure and composition to understand how ITD and height measurement accuracy vary with method, input parameters and data, and stand density. There was little variability in accuracy between the ITD methods where the average F-score and standard deviation (±SD) were 0.47 ± 0.03 using a lower pulse density ALS dataset with an average of 8 pulses per square meter (ppm2) and 0.50 ± 0.02 using a higher pulse density ALS dataset with an average of 22 ppm2. Using higher ALS pulse density data produced higher ITD accuracies (F-score increase of 10–13%) in some of the methods versus more modest gains in other methods (F-score increase of 1–3%). Omission errors were strongly related with stand density and largely consisted of suppressed trees underneath the dominant canopy. Simple canopy height model (CHM)-based methods that utilized fixed-size local maximum filters had the lowest omission errors for trees across all canopy positions. ITD accuracy had large intra-method variation depending on input parameters; however, the highest accuracies were obtained when parameters such as search window size and spacing thresholds were equal to or less than the average crown diameter of trees in the study area. All ITD methods produced height measurements for the detected trees that had low RMSE (<1.1 m) and bias (<0.5 m). Overall, the results from this study may help guide end-users with ITD method application and highlight future ITD method improvements.

Recognition of specified errors of Individual Tree Detection methods based on Canopy Height Model

In the last two decades, many individual tree detection (ITD) methods have been developed based on Airborne Laser Scanner (ALS) data, but their application still has some shortcomings. In complex and dense forest stands, the number of tree segments extracted from Canopy Height Model (CHM) are frequently over- and under-estimated, reducing the ability to predict stand structure. This study presents the possibility of distinguishing correct from erroneous segments resulting from the ITD methods based on CHM. For this purpose, three machine learning methods were tested: Random Forest (RF), Support Vector Machine and k-Nearest Neighbor among which the RF algorithm gave the best results. Groups of predictors based on segment geometry as well as structural and intensity metrics from the ALS point cloud were used. From the whole set of predictors, the ratio of segment perimeter to its area proved to be the most important. Using RF classifier, it was possible to identify under-segmentation and over-segmentation errors, as well as correct segments, with high accuracies for training (OA = 87.0% and κ = 0.794) and test data (OA = 85.3 and κ = 0.641). Recognition of specific segmentation errors is important in that it can be used to determine which conditions favour the occurrence of errors and which tree species are more likely to be incorrectly segmented. Therefore, the study verified the error susceptibility of individual tree species. This can help improve calibration and development of segmentation algorithms according to the tree species analysed.

Parameterization of the Individual Tree Detection Method Using Large Dataset from Ground Sample Plots and Airborne Laser Scanning for Stands Inventory in Coniferous Forest

Highly accurate and extensive datasets are needed for the practical implementation of precision forestry as a method of forest ecosystem management. Proper processing of huge datasets involves the necessity of the appropriate selection of methods for their analysis and optimization. In this paper, we propose a concept for and implementation of a data preprocessing algorithm, and a method for the empirical verification of selected individual tree detection (ITD) algorithms, based on Airborne Laser Scanning (ALS) data. In our study, we used ALS data and very extensive dendrometric field measurements (including over 21,000 trees on 522 circular sample plots) in the economic and protective coniferous stands of north-eastern Poland. Our algorithm deals well with the overestimation problems of tree top detection. Furthermore, we analyzed segmentation parameters for the two currently dominant ITD methods: Watershed (WS) and Local Maximum Filter with Growing Region (LMF+GR). We optimized them with respect to minimizing the Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE). Additionally, our results show the crucial importance of the quality of empirical data for the correct evaluation of the accuracy of ITD algorithms.

Assessment of an improved individual tree detection method based on local-maximum algorithm from unmanned aerial vehicle RGB imagery in overlapping canopy mountain forests

ABSTRACT Low consumer-grade cameras attached to small unmanned aerial vehicles (UAV) can easily acquire high spatial resolution images, leading to convenient forest monitoring at small-scales for forest managers. However, most studies were carried out in the low canopy density and flat ground plantations to detect individual trees. We selected overlapping canopy plantation in mountainous area in the eastern of China and acquired high spatial resolution UAV RGB images to detect individual trees. A total of 402 reference trees were located in three rectangle plots (900 m2). To enhance the confidence of the tested individual tree detection method, clear-cutting and Real-Time Kinematic (RTK) were used to obtain the truth values in the plots. A novel method for semi-automatic individual tree detection was proposed based on a local-maximum algorithm and UAV-derived DSM data (LAD) in this study. The detection accuracy of LAD was compared with commonly used methods based on UAV-derived orthophoto images, local-maximum algorithm (LAO), object-oriented feature segmentation (OFS), multiscale segmentation technique (MST) and manual visual interpretation (MVI). The overall accuracy (OA (%) decreased in the order of LAD (84.5%) > MST (69.1%) > OFS (65.1%) > MVI (64.1%) > LAO (59.1%). LAD had only 15.5%s omission errors (OM (%), which was less than half of the other four methods in comparison. It was noteworthy that MVI had 35.9% OM %, which revealed that MVI should be used carefully as the truth value. LAD showed similar repeated detection error (RP (%) and completely wrong detection (CW (%), while the other four methods had obviously higher CW % than the RP %. From our results, it can be concluded that the proposed LAD method may help improving the accuracy of individual tree detection to an acceptable accuracy (>80%) in dense mountain forests, and has practical advantages in future research direction to assess tree attributes from UAV RGB image.

Model-Based Estimation of Forest Inventory Attributes Using Lidar: A Comparison of the Area-Based and Semi-Individual Tree Crown Approaches

The use of individual tree detection methods to support forest management inventories has been a research topic for over two decades, but a formal assessment of these methods to produce stand-level and region-level predictions of forest attributes and measures of error is lacking. We employed model-based estimation methods in conjunction with the semi-individual tree crown approach (s-ITC) to produce predictions and measures of error for tree volume (VOL), basal area (BA), stem density (DEN), and quadratic mean diameter (QMD) at the scale of forest stands and the entire study region. We compared the s-ITC approach against the area-based approach (ABA) for predictions of region-level and stand-level attributes via model-based root mean squared errors (RMSEs). The study was conducted at the Panther Creek watershed in Oregon, USA using a set of 78 field plots and aerial lidar information. For region-level attributes, s-ITC RMSEs demonstrated changes between −31% and 17% relative to ABA models. At the stand level, median s-ITC RMSEs generally increased, with changes between −29% and 414% relative to ABA models, but demonstrated important reductions in stands where segmentation provided large increases in sample size and was less prone to extrapolation than ABA models. The ABA demonstrated smaller RMSEs in stands without sampled population units for all variables. Our findings motivate further research into niche applications where s-ITC models may consistently outperform ABA models.

Mapping individual trees with airborne laser scanning data in an European lowland forest using a self-calibration algorithm

Traditional field-based forest inventories tend to be expensive, time-consuming, and cover only a limited area of a forested region. Remote sensing (RS), especially airborne laser scanning (ALS) has opened new possibilities for operational forest inventories, particularly at the single-tree level, and in the prediction of single-tree characteristics. Throughout the world, forests have varying characteristics that necessitate the development of modern, effective, and versatile tools for ALS data processing. To address this need, we aimed to develop a tool for individual tree detection (ITD) utilising a self-calibrating algorithm procedure and to verify its accuracy using the complicated forest structure of near natural forests in the temperate zone.This study was carried out in the Polish part of the Białowieża Forest (BF). The airborne laser scanner (ALS) and color-infrared (CIR) datasets were acquired for more than 60 000 ha. Field-based measurements were performed to provide reference data at the single tree level. We introduced a novel ITD method that is self-calibrated and uses a hierarchical analyses of the canopy height model.There were more than 20 000 000 of trees in first layer in BF above 7 m height. Trees visible from above were divided into coniferous, deciduous and mixed trees that were then matched with an accuracy of 85 %, 85 % and 75 %, respectively. Compared to existing methods, the proposed method is more flexible and achieves better results, especially for deciduous species. Before application of the presented method to other regions, the calibration based on the developed optimisation procedure is needed.

Multi-layered tree crown extraction from LiDAR data using graph-based segmentation

With the development of Light Detection and Ranging (LiDAR) and Unmanned Aerial Vehicle (UAV) technology, extracting tree crowns from LiDAR data and infering their geometrical features are becoming more available for everyone. Although the forest has a significant stratification phenomenon in the vertical direction, the existing individual tree detection methods generally aimed at solving overstory trees detection. As a result, the understory trees cannot be effectively extracted. To effectively detect individual tree from multi-layer forests, a multi-layered tree extraction method using a graph-based segmentation algorithm was proposed in this study. First, using the graph-based segmentation algorithm delineates the canopy of the overstory tree on the canopy height model (CHM) generated by LiDAR data. Then, using the sliding window detection method extracts the LiDAR data of understory trees. Finally, the information of understory trees is extracted by the graph-based segmentation algorithm. To verify the performance of the proposed method, this study selected six experimental plots from two research areas. According to the result of our method, the highest matching score and average score for overstory trees reached 91.3 and 86.3; the highest matching score and average score for understory trees reached 78.1 and 63.2. Compared with other methods, our method has better detection results. The experimental results show that the proposed method can extract the understory and overstory trees effectively, thus improving the accuracy of individual tree extraction in multi-layered forests.

Using Tree Detection Based on Airborne Laser Scanning to Improve Forest Inventory Considering Edge Effects and the Co-Registration Factor

The estimation of forest biophysical attributes improves when airborne laser scanning (ALS) is integrated. Individual tree detection methods (ITD) and traditional area-based approaches (ABA) are the two main alternatives in ALS-based forest inventory. This study evaluated the performance of the enhanced area-based approach (EABA), an edge-correction method based on ALS data that combines ITD and ABA, at improving the estimation of forest biophysical attributes, while testing its efficiency when considering co-registration errors that bias remotely sensed predictor variables. The study was developed based on a stone pine forest (Pinus pinea L.) in Central Spain, in which tree spacing and scanning conditions were optimal for the ITD approach. Regression modeling was used to select the optimal predictor variables to estimate forest biophysical attributes. The accuracy of the models improved when using EABA, despite the low-density of the ALS data. The relative mean improvement of EABA in terms of root mean squared error was 15.2%, 17.3%, and 7.2% for growing stock volume, stand basal area, and dominant height, respectively. The impact of co-registration errors in the models was clear in the ABA, while the effect was minor and mitigated under EABA. The implementation of EABA can highly contribute to improve modern forest inventory applications.

Effect of canopy structure on the performance of tree mapping methods in urban parks

Urban forests consist of patches of recreational areas, parks, and single trees on roadsides and other forested urban areas. Large number of tree species and heterogeneous growing conditions result in diverse canopy structure. High variation can be found both at level of single tree crowns and in canopy characteristics of larger areas. As urban forests are typically managed with small-scale, even tree-level operations, there is a need for detailed forest information. In this study, the effect of varying canopy conditions was tested on nine individual tree detection (ITD) methods. All methods utilized airborne laser scanning (ALS)-derived canopy height models (CHM) and different modifications of watershed segmentation (WS).The performance of mapping methods was compared in three strata with varying mean height and canopy cover. The results showed considerable variation between the methods when tested in varying canopy conditions. Especially, presence of large broadleaved trees affected the accuracy of detecting individual trees. The best performing methods for the three strata were G0.7, F2 and Gadapt. The areas with low canopy cover turned out problematic for all ITD methods tested as co-occurrence of small trees and large deciduous trees affected the accuracy significantly. Overall.The results show that stratification can be used to enhance the quality of ITD in urban park areas. However, heterogeneous canopy structure and varying growth patterns typical for urban parks lower the accuracy of tree detection. Also, according to our results we suggest that canopy height and canopy cover alone are insufficient attributes for stratifying urban canopy conditions.

Automated detection of individual clove trees for yield quantification in northeastern Madagascar based on multi-spectral satellite data

There is an increasing demand for clove products, mainly dried buds and essential oil on global markets. Consequently, the importance of clove trees as a provisioning service is increasing at the local level, particularly for smallholders cultivating clove trees as cash crops. Due to limited availability of data on local production, using remote sensing-based methods to quantify today's clove production is of key interest. We estimated the clove bud yield in a study site in northeastern Madagascar by detecting individual clove trees and determining relevant production systems, including pasture and clove, clove plantation and agroforestry systems. We implemented an individual tree detection method based on two machine learning approaches. Specifically, we proposed using a circular Hough transform (CHT) for the automated detection of individual clove trees. Subsequently, we implemented a tree species classification method using a random forests (RF) classifier based on a set of features extracted for relevant trees in the above production systems. Finally, we classified and mapped different production systems. Based on the number of detected clove trees growing in a clove production system, we estimated the production system-dependent clove bud yield. Our results show that 97.9% of all reference clove trees were detected using a CHT. Classifying clove and non-clove trees resulted in a producer accuracy of 70.7% and a user accuracy of 59.2% for clove trees. The classification of the clove production systems resulted in an overall accuracy of 77.9%. By averaging different clove tree yield estimates obtained from the literature, we estimated an average total yield of approximately 575 tons/year for our 25,600 ha study area. With this approach, we demonstrate a first step towards large-scale clove bud yield estimation using remote sensing data and methodologies.

Quantification of uncertainty in aboveground biomass estimates derived from small-footprint airborne LiDAR

To address uncertainty in biomass estimates across spatial scales, we determined aboveground biomass (AGB) in Californian forests through the use of individual tree detection methods applied to small-footprint airborne LiDAR. We propagated errors originating from a generalized allometric equation, LiDAR measurements, and individual tree detection algorithms to AGB estimates at the tree and plot levels. Larger uncertainties than previously reported at both tree and plot levels were found when AGB was derived from remote sensing. On average, per-tree AGB error was 135% of the estimated AGB, and per-plot error was 214% of the estimated AGB. We found that from tree to plot level, the allometric equation constituted the largest proportion of the total AGB uncertainty. The proportion of the uncertainty associated with remote sensing errors was larger in lower AGB forests, and it decreased as AGB increased. The framework in which we performed the error propagation analysis can be used to address AGB uncertainties in other ecosystems and can be integrated with other analytical techniques.

Estimating urban above ground biomass with multi-scale LiDAR

BackgroundUrban trees have long been valued for providing ecosystem services (mitigation of the “heat island” effect, suppression of air pollution, etc.); more recently the potential of urban forests to store significant above ground biomass (AGB) has also be recognised. However, urban areas pose particular challenges when assessing AGB due to plasticity of tree form, high species diversity as well as heterogeneous and complex land cover. Remote sensing, in particular light detection and ranging (LiDAR), provide a unique opportunity to assess urban AGB by directly measuring tree structure. In this study, terrestrial LiDAR measurements were used to derive new allometry for the London Borough of Camden, that incorporates the wide range of tree structures typical of an urban setting. Using a wall-to-wall airborne LiDAR dataset, individual trees were then identified across the Borough with a new individual tree detection (ITD) method. The new allometry was subsequently applied to the identified trees, generating a Borough-wide estimate of AGB.ResultsCamden has an estimated median AGB density of 51.6 Mg ha–1 where maximum AGB density is found in pockets of woodland; terrestrial LiDAR-derived AGB estimates suggest these areas are comparable to temperate and tropical forest. Multiple linear regression of terrestrial LiDAR-derived maximum height and projected crown area explained 93% of variance in tree volume, highlighting the utility of these metrics to characterise diverse tree structure. Locally derived allometry provided accurate estimates of tree volume whereas a Borough-wide allometry tended to overestimate AGB in woodland areas. The new ITD method successfully identified individual trees; however, AGB was underestimated by ≤ 25% when compared to terrestrial LiDAR, owing to the inability of ITD to resolve crown overlap. A Monte Carlo uncertainty analysis identified assigning wood density values as the largest source of uncertainty when estimating AGB.ConclusionOver the coming century global populations are predicted to become increasingly urbanised, leading to an unprecedented expansion of urban land cover. Urban areas will become more important as carbon sinks and effective tools to assess carbon densities in these areas are therefore required. Using multi-scale LiDAR presents an opportunity to achieve this, providing a spatially explicit map of urban forest structure and AGB.

DETECTING FORESTS DAMAGED BY PINE WILT DISEASE AT THE INDIVIDUAL TREE LEVEL USING AIRBORNE LASER DATA AND WORLDVIEW-2/3 IMAGES OVER TWO SEASONS

Abstract. Pine wilt disease is caused by the pine wood nematode (Bursaphelenchus xylophilus) and Japanese pine sawyer (Monochamus alternatus). This study attempted to detect damaged pine trees at different levels using a combination of airborne laser scanning (ALS) data and high-resolution space-borne images. A canopy height model with a resolution of 50 cm derived from the ALS data was used for the delineation of tree crowns using the Individual Tree Detection method. Two pan-sharpened images were established using the ortho-rectified images. Next, we analyzed two kinds of intensity-hue-saturation (IHS) images and 18 remote sensing indices (RSI) derived from the pan-sharpened images. The mean and standard deviation of the 2 IHS images, 18 RSI, and 8 bands of the WV-2 and WV-3 images were extracted for each tree crown and were used to classify tree crowns using a support vector machine classifier. Individual tree crowns were assigned to one of nine classes: bare ground, Larix kaempferi, Cryptomeria japonica, Chamaecyparis obtusa, broadleaved trees, healthy pines, and damaged pines at slight, moderate, and heavy levels. The accuracy of the classifications using the WV-2 images ranged from 76.5 to 99.6 %, with an overall accuracy of 98.5 %. However, the accuracy of the classifications using the WV-3 images ranged from 40.4 to 95.4 %, with an overall accuracy of 72 %, which suggests poorer accuracy compared to those classes derived from the WV-2 images. This is because the WV-3 images were acquired in October 2016 from an area with low sun, at a low altitude.

Stand density estimators based on individual tree detection and stochastic geometry

Individual tree detection methods leave smaller trees hiding below larger trees undetected. This is a problem for remote sensing forest inventories, leading, for example, to severe underestimation ...

International Benchmarking of the Individual Tree Detection Methods for Modeling 3-D Canopy Structure for Silviculture and Forest Ecology Using Airborne Laser Scanning

Canopy structure plays an essential role in biophysical activities in forest environments. However, quantitative descriptions of a 3-D canopy structure are extremely difficult because of the complexity and heterogeneity of forest systems. Airborne laser scanning (ALS) provides an opportunity to automatically measure a 3-D canopy structure in large areas. Compared with other point cloud technologies such as the image-based Structure from Motion, the power of ALS lies in its ability to penetrate canopies and depict subordinate trees. However, such capabilities have been poorly explored so far. In this paper, the potential of ALS-based approaches in depicting a 3-D canopy structure is explored in detail through an international benchmarking of five recently developed ALS-based individual tree detection (ITD) methods. For the first time, the results of the ITD methods are evaluated for each of four crown classes, i.e., dominant, codominant, intermediate, and suppressed trees, which provides insight toward understanding the current status of depicting a 3-D canopy structure using ITD methods, particularly with respect to their performances, potential, and challenges. This benchmarking study revealed that the canopy structure plays a considerable role in the detection accuracy of ITD methods, and its influence is even greater than that of the tree species as well as the species composition in a stand. The study also reveals the importance of utilizing the point cloud data for the detection of intermediate and suppressed trees. Different from what has been reported in previous studies, point density was found to be a highly influential factor in the performance of the methods that use point cloud data. Greater efforts should be invested in the point-based or hybrid ITD approaches to model the 3-D canopy structure and to further explore the potential of high-density and multiwavelengths ALS data.

Individual Tree Detection from High Spatial Resolution Imagery Using Color and Texture Features

An automatic individual tree detection method from pure image is proposed. Color and texture features are selected to form a vector for a pixel-level classification, then trained to assign a label to each pixel. Other features can be integrated into the pixel vector for extracting more information of trees. An ensemble method combining multiple logistic regression classifiers improves effectiveness of the single pixel-level classifier. Then spectral, shape and knowledge characteristics of individual tree crowns are used for tree top localization. At last, tree crowns are delineated by region-based algorithm.

Individual Tree Detection from High Spatial Resolution Imagery Using Color and Texture Features

An automatic individual tree detection method from pure image is proposed. Color and texture features are selected to form a vector for a pixel-level classification, then trained to assign a label to each pixel. Other features can be integrated into the pixel vector for extracting more information of trees. An ensemble method combining multiple logistic regression classifiers improves effectiveness of the single pixel-level classifier. Then spectral, shape and knowledge characteristics of individual tree crowns are used for tree top localization. At last, tree crowns are delineated by region-based algorithm.